JP3168529B2 - Load prediction method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a load estimating method for estimating a periodically fluctuating heat load or electric power load in a facility such as a heat source plant or a factory.

[0002]

2. Description of the Related Art In a heat source plant or the like, a heat storage system for effectively utilizing nighttime power or the like is becoming popular. In such a system, in order to efficiently operate the heat source device, prediction of the heat load and the power load is extremely important.

[0003] As a method for predicting this load, literature: Box & Jenkins: Time Seri
es Analysis--forcasting control-, Holden-day (197
0)] ARMA (autoregressive moving average)
A load prediction method using an ARIMA model formula in which a one-day periodic fluctuation pattern model formula is added to the model formula has been proposed (Japanese Patent Application No. 1-180605). However, according to the load prediction method using such an ARIMA model formula, since the prediction accuracy of a high load (peak load level) predicted at the time before the load rises is affected by the peak load level of the previous day, the prediction accuracy is high. However, there is a problem that the temperature is reduced. Therefore, a method has been proposed in which the predicted load obtained by such a load prediction method is corrected based on external environment prediction information such as the predicted maximum outside temperature of the next day announced by a weather station (Japanese Patent Application No. Hei 3 (1990) -319). No. 36655).

[0004]

In the conventional load prediction method based on the ARIMA model formula, the prediction accuracy of a high load (peak load level) predicted at the time before the load rises is affected by the peak load level of the previous day. For this reason, a good prediction can be made in a season with a small climate change, but there is a problem that the prediction accuracy cannot be guaranteed in a season with a large climate change. In addition, the method of correcting the prediction load based on the ARIMA model formula based on the external environment prediction information has a problem that it is difficult to automate the acquisition of the prediction information such as the predicted maximum outside temperature. Since the correction coefficient is obtained, there is a problem that it is difficult to associate the correction coefficient with the maximum load value. The present invention has been made to solve the above-described problem, and has as its object to provide a load prediction method capable of performing load prediction with high accuracy without receiving prediction information from the outside.

[0005]

SUMMARY OF THE INVENTION According to the present invention, a new case is obtained by phase case-based modeling in which actual measured data at the current time inputted as a new case is predicted using a case base using past measured data as a case. It is characterized in that the maximum load of the day is predicted from the case and the predicted load by the ARIMA model formula is corrected based on the maximum load. In addition, the prediction of the maximum load by phase case-based modeling is based on creating a monthly case base from past measured data and learning when the measured data is newly obtained, updating the case base, and When input, a case similar to this new case is extracted from the case base, the importance of the similar case at the time of load inference is determined based on the distance between the new case and the extracted similar case, and based on this importance, The maximum load is inferred from similar cases on a case basis for each month, and a membership function that indicates the membership of each month is created from past measured data, and the membership of the new case is assigned to each month from this membership function. Determine gender,
The maximum load obtained for each month's case base is integrated into one based on the attribution, and the final maximum load is obtained, using only measurable actual data obtained up to the current time. In which the maximum load is predicted.

[0006]

According to the present invention, the maximum load is predicted from the actual measurement data at the current time as a new case, and the predicted load by the ARIMA model formula is corrected based on the maximum load. In addition, when a new case is input, each month it is necessary to extract cases similar to this new case from the case base, determine the importance of similar cases, and infer the maximum load from similar cases based on this importance. Is performed for each case base, the membership of the new case is determined for each month from the membership function, and the maximum load determined for each case base for each month is integrated into one based on the membership to obtain the final Maximum load is obtained.

[0007]

FIG. 1 is a flowchart for explaining a load estimation method according to an embodiment of the present invention. In the present embodiment, for example, an air conditioning load (heat load, hereinafter abbreviated as load) pattern in a building is predicted,
The peak load of the day is predicted by applying the phase case-based modeling technique proposed in -269129, and the load pattern is corrected using the maximum load.

[0008] First, the phase case based modeling will be described. Topological case-based modeling is a technique for modeling a nonlinear system, and uses a case base in which past history data is accumulated as cases. In general case-based inference, inference is performed based on the similarity between cases, but there is no general definition for this similarity determination method. In this method, a system in which continuous mapping from an input space to an output space is established is applied, and similarity determination is performed based on the phase of the input space. The phase referred to here is a concept of proximity in one space.
Since the points are guaranteed to be close even in the output space, this phase can be used for determining the similarity between cases.

In the modeling in the phase case-based modeling, input / output history data, which is actually measured data, is collected from an actual air conditioning system to determine a model structure, and the past history data is divided into months and collected in January. A case base for January is created based on the collected data, and a case base is created for each month such as a case base for December based on data collected in December. The reason for creating a monthly basis in this way is that if an attempt is made to create a case base in an annual range, it becomes very complicated. By limiting the range to a monthly range, the accuracy of the model can be improved.

In the creation of such a case base, first, an input space is quantized and divided into a finite number of input events, and an input / output history data belonging to the same input event is integrated to create one case. Note that the same input event is, for example, when the input variable of the outside temperature is quantized every 1 ° C., 25.2 ° C., 2
This is historical data such as 5.4 ° C.

Then, the quantized input variable value is expressed as ｛X
i｝ (i = 1, 2,..., n: Xi is, for example, the outside air temperature,
If the number of pieces of history data belonging to the same input event is p, the output average value Y is the average value of the output history data (the highest load in this embodiment) belonging to the same input event. Is given by

[0012]

(Equation 1)

YDb (b = 1, 2,...,
p) is output history data belonging to the same input event. Thus, one case of the output value Y for the input variable value {Xi}, that is, the case of {Xi} → Y has been created. Since the same input event number p differs for each case, p is stored in the case base for each case. In this way, the history data is collected every month, and a case base is created every month.

Next, the created case-based learning is performed as follows. In a certain month, when new history data such as input history data {XDinew} and output history data YDnew is obtained, a case having the same input event as this input history data is extracted from the case base of the month, and To update.

Assuming that the extracted case is {Xi} → Yold and the number of identical input events of the extracted case is p, the output value Ynew of the revised case is calculated as follows, and p is increased by 1 to p
+1. Ynew = (pαYold + YDnew) / (pα + 1) (2)

The learning parameter α in the equation (2) is determined as follows. First, the output value change amount β is calculated from the output value Yold of the extracted case and the new output history data YDnew as in the following equation. β = (| Yold−YDnew |) / Yold (3) As shown in FIG. 2, the learning parameter α is inversely proportional to the amount of change based on the normal aging of the target system. The learning parameter α is determined by applying the quantity β to the relationship shown in FIG.

Then, based on the learning parameter α,
The output value Ynew of the revised case is obtained using the equation (2). Thus, the case of {Xi} → Yold becomes {Xi}
→ Revise to the case of Ynew.

The learning using the learning parameter α has the following meaning. A large amount of change β, that is, a large difference between the output value Yold of the past case and the new output history data YDnew means that the system is rapidly changing over time (for example, population increase). Therefore, when a sudden change occurs, the learning parameter α is reduced so that the influence of the past case is reduced. As a result, it is possible to provide adaptability that can follow rapid changes in the system.

Next, in preparation for determining the membership which indicates to which month the new case belongs to at the time of output inference, which will be described later, a membership function representing the membership is input to the input variable ｛X.
It is created in advance for every i｝ (i = 1, 2,..., n). To create this membership function, the annual range is first quantized and divided for each input variable as shown in FIG.

Next, using the historical data for each month when the case base was created, a frequency distribution indicating where and how many of the historical data belongs to the above-mentioned annual range is created for each month. FIG. 4 shows such a frequency distribution for month f out of January to December. Then, the frequency distribution is normalized by the maximum value fmax in month f to generate a possibility distribution, which is used as the membership function for month f as shown in FIG. Thus, the membership functions for January to December are created as shown in FIG.

Such membership functions are created for each input variable {Xi}, and the membership of the new case to each month is recognized using the created n types of membership functions. Of the ship functions, those whose membership function's center of gravity for each month is evenly distributed over the annual range, that is, those that show the characteristics of each month and are easy to recognize, are used for membership recognition. I do. Here, the input variables of the adopted membership function are assumed to be state recognition variables {Xj} (j = 1,..., K: k ≦ n). That is, k kinds of membership functions are prepared.

Next, when new input data is input as a new case, an output value corresponding to this input data is inferred using the case base created and learned as described above. Since the inference of the output value proceeds based on the similarity between cases, the similarity will be described first.

The similarity between the new case and the existing case in the case base is determined by the neighborhood of the input space, and the neighborhood of the input space is determined from the neighborhood of the output space. That is, in the vicinity of the input space, as shown in FIGS. 7A and 7B, cases (black circles in the figure) included in the neighborhood ε of the output space Y determined by the required inference error are collected. These are obtained as distributions when they are projected onto each variable axis of the input space {Xi} (FIG. 7 shows only X1 and X2 as an example).

Thus, the neighborhood {δi} (i =
1, 2,..., N), there are actually a plurality of neighborhoods ε in the output space, and neighborhoods ε in each output space.
, The neighborhood {δi} of the input space is determined.
.., δn are obtained by statistical processing for each i｝.

Based on the neighborhood {δi} in the input space, the similarity between the new case and the existing case in the case base is defined as follows. In the existing case having the highest similarity of 0, | Xi * −Xi | = 0 holds for all i. Here, {Xi *} (i = 1, 2,...)
.., N) are input variable values of the new case, and these input variable values are assumed to be discretized similarly to the input variable values {Xi} of the existing case.

Further, the existing case of the similarity Q (Q> 0) is as follows:
| Xi * −Xi | ≦ δi + Q−1 holds for all i, and at least one i that satisfies | Xi * −Xi | = δi + Q−1 exists. Based on such similarity, existing cases similar to the new case are extracted from the case base in descending order of similarity, and these are regarded as similar cases.
At this time, the number m of existing cases to be extracted is m = n + 1 with respect to the dimension n of the input space.

Next, the importance of the extracted similar case at the time of output value inference is calculated based on the distance between the new case and the similar case. In the present embodiment, the new case and the c-th case (c = 1,.
, M), the distance Lc to the similar case is defined as follows.

[0028]

(Equation 2)

In the equation (4), Xic is the input variable value of the c-th similar case, and Ric is the degree of influence of the input variable Xic on the output Y, that is, the correlation coefficient. Then, the importance Wc of the c-th similar case is defined as follows. Wc = exp (−Lc) (5)

Next, using the extracted m similar cases, an output inference value Y * corresponding to the input variable value {Xi *} of the new case is obtained by the following equation.

[0031]

(Equation 3)

Here, Yc is the output value of the c-th similar case. Thus, an output inference value Y * is obtained. The extraction of similar cases, the calculation of the distance and importance, and the inference of the output inference value Y * as described above are calculated using each case base from January to December. As a result, 12 output inference values Y are obtained. * Will be obtained.

Next, the output inference value Y * obtained for each case base of each month is integrated using a month recognition method based on fuzzy logic to obtain a final output inference value. For this purpose, the membership of the new case to each month is determined using the k types of membership functions described above.

First, from the input variable values {Xi *} of the new case, the state recognition variables {Xj} (j = 1,.
Extract {Xj *} corresponding to k). This is, for example, two of the outside air temperature X1 and the outside air humidity X2 as state recognition variables.
If one (k = 2) is adopted, it means that the input variable values of the new case of the outside temperature X1 * and the outside air humidity X2 * are taken out.

Then, in the membership function of the state recognition variable Xj as shown in FIG. 6 (however, Xi in FIG. 6),
If the attribution to f month is Af (Xj), the attribution Af to f month of the extracted new case {Xj *} (j = 1,..., k) can be calculated as follows.

[0036]

(Equation 4)

Thus, the attribution A of the new case to each month A
f (f = 1,..., 12) is determined. The final output inference value Y * is, assuming that the output inference value on a case basis for each month obtained above is {Yf *} (f = 1,..., 12), these output inference values {Yf *} Is obtained by the following equation by the load average according to the attribute Af of

[0038]

(Equation 5)

Thus, a final output inference value is obtained.
The present invention predicts the maximum load of the day as an output inference value using the above-described phase case-based modeling technology. In this embodiment, the actually measured outside air temperature and outside air humidity are set as input history data {XDi} (the dimension n of the input variable is n = 2), and the corresponding maximum load actual measurement value is output history data YD. Then, a case base of {Xi} → Y is created from these history data.

With respect to this case base, if the outside air temperature and the outside air humidity (actually measured values) at the current time of the day are input as the input variable values {Xi *} of the new case, the predicted maximum load value is obtained as the output inference value Y *. Obtainable.

Next, a load prediction method according to the present invention will be described with reference to FIG.
This will be described with reference to FIG. First, it is determined whether or not the day has begun (step 100), and if so, the correction execution flag is set to a set state (for example, “1”) (step 1).
01). Next, the outside air temperature and the outside air humidity at the current time T are actually measured to obtain input variable values {Xi *} of the new case, and the maximum load predicted value Tc is obtained by using the above-described phase case base modeling (step 102). In this embodiment, the time and the time are set on the assumption that the prediction is performed every hour. Also, since the load prediction starts at 0:00, the current time is T
= 0.

Subsequently, a load pattern is predicted using an ARIMA model formula obtained by adding a one-day periodic fluctuation pattern model formula to the ARMA model formula (step 103).
Since the ARIMA model formula is described in detail in Japanese Patent Application No. 1-180605, the description thereof is omitted here.

The load prediction based on the ARIMA model formula uses the measured load data from the current time to 24 hours before the current
By substituting into the IMA model equation, the load is predicted one hour ahead (the direction in which time advances is first, and the direction in which time returns is front. Therefore, here, 1:00). Then, the obtained predicted load one hour ahead and the measured load data from the current time to 23 hours before are obtained by ARIMA.
By substituting into the model formula, the load 2 hours ahead (2 o'clock) is predicted. Similarly, from the predicted load one or two hours ahead and the current time, 22
The load data up to the previous hour is substituted into the ARIMA model formula, and the load 3 hours ahead (3 o'clock) is predicted. Hereinafter, the load from the current time to 24:00 is predicted by this repetition, and a load pattern like I in FIG. 8 is obtained.

Next, whether or not to correct the load pattern based on the ARIMA model formula is determined based on the state of a correction execution flag described later (step 104). If the correction execution flag is set, the determination is Y.
The answer is es and the routine proceeds to step 105, where it is determined whether or not it is after the determination start time. This determination start time is set at 8:00 in this embodiment.

Since the current time is T = 0, the determination is No, and it is next determined whether or not the maximum load predicted value Tc obtained in step 102 is present (step 109). this is,
If the maximum load prediction fails, there is no maximum load prediction value Tc,
This is because correction is impossible, and if there is no predicted value Tc, the processing at the current time is ended without performing the correction described below, using the load pattern based on the ARIMA model formula as the final load pattern (step 113). ).

If there is the maximum predicted load value Tc, the minimum value BASET in the load pattern from time T + 1 to 24:00 according to the measured load data from time 0 to time T and the ARIMA model formula is obtained (step 110). Where T
The value +1 indicates a time one prediction processing interval ahead of the time T. In this embodiment, since the processing is performed every hour, it is a time one hour later.

The time T according to the ARIMA model formula
The highest value ARTmax from the load patterns from +1 to 24:00
Is obtained (step 111). Next, the predicted load ART (t) at time t (t = T + 1 to 24) based on the ARIMA model obtained at the current time T is corrected, and the corrected predicted load * ART is obtained.
(T) is obtained as in the following equation. * ART (t) = C {ART (t) -BASET} + BASET (9)

Here, C is a correction coefficient, which is given by the following equation. C = (Tc-BASET) / (ARTmax-BASET) (10) Such a corrected predicted load * ART (t) is expressed as t = T + 1 to 1.
24 are sequentially obtained. With this corrected predicted load, the load pattern can be corrected as shown in II of FIG. 8 (step 112).

Thus, the processing at the current time T = 0 is completed. Next, when the current time becomes T = 1, the outside air temperature and the outside air humidity are measured, and the load prediction and the correction are performed in the same manner as described above. . Similarly, load prediction and correction are performed as time elapses as T = 2, 3,.... However, during T = 1 to 5:00, the maximum load prediction value Tc obtained at T = 0 is used, and the maximum load prediction in step 102 is not performed.

Next, when the current time reaches T = 6: 00, the maximum load prediction value Tc is obtained by executing the maximum load prediction in step 102 again, and the other load prediction and correction are performed in the same manner as described above. When T = 7, the maximum load predicted value Tc obtained at 6:00 is used. Further, when the current time reaches T = 8: 00, the maximum load prediction in step 102 is executed again to obtain the maximum load predicted value Tc, and this value is used at times thereafter on that day.

Further, in this embodiment, since the above-mentioned judgment start time is set to 8:00, the result of step 105 becomes Yes because T = 8: 00, and the correction stop condition shown in the following equation is calculated (step 106). . | ART-1 (T) -act (T) | ≦ | * ART-1 (T) -act (T) | (11)

In equation (11), ART-1 (T) is equal to T
ARIMA obtained at -1 o'clock, one hour before (7 o'clock)
Predicted load at the current time T (8 o'clock) by model, similarly * A
RT-1 (T) is the corrected predicted load at the current time T obtained at the time T-1 and act (T) is the measured load value at the current time T. Then, it is determined whether or not Expression (11) is satisfied (Step 107). If it is, the correction execution flag is reset (for example, “0”) (Step 10).
8).

If the equation (11) does not hold, the routine proceeds to step 109, where the load pattern is corrected in the same manner as described above. Thereafter, the correction stop determination is performed every hour until Expression (11) is satisfied.

The correction stop determination is performed for the following reason. That is, each time the current time elapses one prediction processing interval (one hour in the embodiment), a new actual load value is obtained from the air conditioning system. The prediction by the ARIMA model formula predicts the load pattern while taking in the actual measured value of the current time. When the load rises, the actual measured value reflecting the influence of climate change is obtained. Approach the actual load that would be obtained. This is a feature of the load prediction method using the ARIMA model formula.

Therefore, if the load rises at the current time of, for example, 8:00 or 9:00, sufficient accuracy can be secured at the time after the load rises only by the prediction using the ARIMA model, and the above-described correction is continued. Then, the correction may be excessive. Therefore, equation (11)
When the difference between the corrected predicted load and the measured load value is equal to or greater than the difference between the predicted load and the measured load value, that is, when the determination result that the predicted load is closer to the measured load value than the corrected predicted load is obtained, the correction is performed. It will not be done.

At the time after the correction execution flag is reset, the determination in step 104 indicates N
Since the result is o, the process proceeds to step 113 and ends the processing without correcting the load pattern.

As described above, the load can be reduced by using the load prediction by the ARIMA model formula having a practical level of performance with respect to the prediction accuracy several hours ahead and the phase case-based modeling capable of coping with nonlinear factors such as climate change. It is possible to improve the accuracy of predicting a high load (peak load) before starting.

[0058]

According to the present invention, the maximum load is predicted directly from the actual measurement data at the current time, which is a new case, and the predicted load based on the ARIMA model formula is corrected based on the maximum load. There is no need to obtain prediction information such as the maximum outside temperature. In addition, since the predicted load to be corrected is corrected using the highest predicted load directly corresponding to the predicted load, sufficient correction accuracy can be obtained.

In the maximum load prediction based on the phase case-based modeling, an optimal similar case is extracted for a new case, the maximum load is predicted, and an appropriate month model is determined from the membership function for the new case. As a result, a final maximum load can be obtained, so that a high-precision maximum load prediction can be performed even at a time when the climate change is large. Accordingly, a highly accurate load prediction can be performed even at a time when the fluctuation of the peak load level is large.

[Brief description of the drawings]

FIG. 1 is a flowchart for explaining a load prediction method according to an embodiment of the present invention.

FIG. 2 is a diagram showing a determination state of learning parameters.

FIG. 3 is a diagram showing an example in which an annual range of an input variable is subdivided.

FIG. 4 is a diagram showing a distribution state of history data for each month.

FIG. 5 is a diagram showing an example in which the distribution of history data is normalized.

FIG. 6 is a diagram showing the distribution of membership functions for each month.

FIG. 7 is a diagram showing the vicinity of an input space.

FIG. 8 is a diagram showing a load pattern and a corrected load pattern obtained by the load prediction method according to the present invention.

[Explanation of symbols]

 I: load pattern, II: correction load pattern.

Claims (2)

(57) [Claims] 1. A load prediction method for predicting an air-conditioning load by an ARIMA model formula in which a one-day periodic fluctuation pattern model formula is added to an ARMA model formula, wherein actual measurement data at the current time inputted as a new case is Predict the maximum load of the day from the new case by phase case-based modeling that makes predictions using a case base using past measured data as a case, and correct the predicted load by the ARIMA model formula based on the maximum load. A load prediction method characterized in that: 2. The load prediction method according to claim 1, wherein in the prediction of the maximum load by the phase case base modeling, a case base for each month is created from past measurement data, and the measurement data is newly obtained. When learning, the case base is updated, and when the new case is input, a case similar to the new case is extracted from the case base, and similarity is determined based on the distance between the new case and the extracted similar case. The importance of the case at the time of load inference is calculated, the highest load is inferred from similar cases based on this importance for each case base of each month, and a membership function indicating the attribution to each month from past measured data. Is determined and the membership of the new case to each month is determined from this membership function, and the maximum load obtained for each case base in each month is integrated into one based on the membership. To is obtained by to obtain the final maximum load, load prediction method characterized by performing the maximum load predicted using only the measurable measurement data obtained up to the current time.